Summary
Epistemic task selection applies the active inference framework to choose what work to do next. Instead of selecting tasks based solely on extrinsic goals (e.g., "need to ship feature X"), it also scores tasks by their epistemic value—the expected reduction in model uncertainty after observing the outcome of doing that task. This creates a formal mechanism for the exploration-exploitation dilemma: the agent continues to gather information (explore) until no further epistemic gains remain, then exploits automatically. The approach was motivated by a concrete software‑engineering self‑model that was data‑starved despite having a complete measurement infrastructure—it needed to actively select observations that would improve its predictive accuracy.
Key Points
- Active inference decomposes expected free energy into pragmatic (extrinsic, goal‑oriented) and epistemic (intrinsic, information‑seeking) components.
- Epistemic value is maximal when the agent can visit states it is currently uncertain about. Once uncertainty is resolved, exploitation takes over.
- Acting is experiment selection: optimal behavior under active inference reduces to Bayesian experiment design when preferences are suppressed (Friston et al., 2021).
- Concretely, tasks (e.g., editing a file in a codebase) are ranked by how uninformative past outcomes have been about them—no graded events, reactive‑vs‑emerging disagreement, or high prediction‑error residuals.
- A lightweight existence proof exists (johanity/theorist): force a guess before each experiment, then use the prediction error as the learning signal.
Concepts
- Pragmatic value: value derived from achieving a preferred outcome (extrinsic reward).
- Epistemic value: value derived from reducing uncertainty about the world (information gain).
- Expected free energy: the total quantity an active‑inference agent minimizes; decomposes into pragmatic and epistemic terms.